Skip to main content
Have a personal or library account? Click to login
An Automated AI-Based Framework for Rapid Progressive Collapse Assessment of Steel Frames Cover

An Automated AI-Based Framework for Rapid Progressive Collapse Assessment of Steel Frames

Open Access
|Nov 2026
Read online

Figures & tables

Table 1:

Comparative assessment of predictive methods for progressive collapse analysis (adapted from Mansour et al., 2025)

Feature [–]DEMFEMAEMPhysics-Based Gaming EnginesML
Modelling detailHighModerateHighHighLow
AccuracyHighHighHighLowHigh
Computational costVery highHighHighLowModerate
Ease of useComplexComplexComplexModerateEasy
Real-time capabilityNoneNoneModerateHighHigh
Figure 1:

Structural bending moments (a) before column removal, and (b) increase in bending moment after column removal due to span increase (reproduced from Mansour et al., 2025)

Figure 2:

VOSviewer author network map of the reviewed literature on AI-based progressive collapse and structural surrogate modelling, with nodes colored by publication year (2023–2026)

Table 2:

Summary of reviewed machine learning studies for progressive collapse and structural surrogate modelling

StudyStructural typologyOutput indicator(s)Dataset scale [–]Key contribution
Fu (2020)Steel framed buildingsCollapse classification (fire induced)Not reported at large scaleFeasibility of ML classification for collapse potential
Zhu et al. (2022)Steel moment framesDIF (dynamic, single indicator)Database of pushdown analysesML prediction of dynamic amplification
Esfandiari et al. (2023)3D reinforced concrete framesCollapse resistance (multi column removal)>600 optimized framesSelf training ML on limited labelled data
Guo et al. (2025)Steel frame composite floorsCollapse resistance and probability (dual)Not directly comparable (mechanics based substructure)Joint prediction with SHAP interpretability
Zhao et al. (2025)General structural frames (static)Displacement>300,000 FE simulationsEnsemble averaging across six surrogate families
Lee et al. (2025a)Structural and thermal response (general)Structural/thermal response fieldsNot progressive collapse specificIdentified data imbalance degradation in DNN surrogates
Mansour et al. (2025) (authors' prior work)Steel moment framesLIF (static, single indicator)3,990 FE modelsFeasibility benchmark of ML vs FE vs DEM/AEM/gaming methods
Present studySteel moment framesLIF and DIF (joint)32,886 configuration–scenario recordsOverfitting-aware benchmarking across 11 architectures; public deployment
Figure 3:

Overall workflow of the proposed framework, from parametric input definition through nonlinear FE analysis, dataset partitioning, and model training, to prediction of the Load Increase Factor and Dynamic Increase Factor

Table 3:

Parametric dataset generation summary

StageCount [–]Description
Base building configurations10,962Factorial combination of stories, story height, span, bays, and column location
Configuration–scenario combinations32,886Base configurations × 3 load scenarios
LIF label source–SAP2000 nonlinear static pushdown analyses
DIF label source–SAP2000 nonlinear dynamic analyses with sudden column removal
Unique records in ML dataset32,886Each labelled jointly with LIF and DIF
Table 4:

Descriptive statistics of the target variables (LIF and DIF) across the full dataset

TargetMinimum [–]Maximum [–]Mean [–]
Load Increase Factor-(LIF)1.023.652.26
Dynamic Increase Factor-(DIF)1.184.242.60
Table 5:

Definition of the three load scenarios

ScenarioLoad pattern on the main beamStructural idealisationApplied inMagnitude / basis
1Uniformly distributed loadOne-way slab actionLIF (nonlinear static pushdown) and DIF (nonlinear dynamic)Base: qDL = 3.0 kN/m2; qLL = 2.5 kN/m2; beam line load = q × tributary width.
2Trapezoidal or triangular distributed loadTwo-way slab action (tributary-area theory)LIF (nonlinear static pushdown) and DIF (nonlinear dynamic)Same base gravity loads; line load varies along the beam according to tributary-area geometry.
3Concentrated load at midspanPoint load at the centre of the main beamLIF (nonlinear static pushdown) and DIF (nonlinear dynamic)Same base gravity load resultant as the affected beam load, represented as a midspan point load.
Figure 4:

Load scenarios applied to the main beam: (a) uniformly distributed load, (b) trapezoidal or triangular distributed load, and (c) concentrated load at midspan

Table 6:

Fitted standardization statistics for the four numerical input features, computed on the training partition

FeatureUnitMeanStandard deviation
Number of stories[–]16.058.36
Story height[m]4.000.81
Bay span[m]6.982.00
Number of bays[–]5.962.58
Table 7:

Dataset partitioning at the configuration level

PartitionConfigurations [–]Records [–]Proportion [%]
Training7,67323,01970
Validation1,0963,28810
Test2,1936,57920
Total10,96232,886100
Table 8:

Benchmarked model architectures and hyperparameters

ModelFamilyKey hyperparameters
Linear regressionClassicalOrdinary least squares
Polynomial regressionClassicalDegree 3
Random forestClassical (ensemble)100 trees
Gradient boostingClassical (ensemble)100 estimators, learning rate 0.1, maximum depth 3
Support-vector regressionClassical (kernel)RBF kernel, C = 10, ɛ = 0.1
XGBoostClassical (ensemble)200 estimators, learning rate 0.05, maximum depth 6
LightGBMClassical (ensemble)200 estimators, learning rate 0.05, 31 leaves per tree
DL-1 SimpleDeep learningFully connected, 64–32, ReLU
DL-2 BN+DropoutDeep learningFully connected, 128–64, ReLU, batch normalisation, dropout
DL-3 DeepDeep learningFully connected, 256–128–64, ReLU
DL-4 RegularisedDeep learningFully connected, 128–128–64, ReLU, batch normalisation, dropout
Table 9:

Effect of the splitting strategy on test accuracy and train–validation gap. Italic: gap above 0.01

ModelSplitTest R2 LIF [–]Test R2 DIF [–]Gap LIF [–]Gap DIF [–]
Random forestRow-level0.97930.97130.01890.0255
Random forestConfiguration0.95890.93770.03910.0574
LightGBMRow-level0.97110.95790.00250.0019
LightGBMConfiguration0.96640.95350.01100.0096
XGBoostRow-level0.96560.95290.00730.0063
XGBoostConfiguration0.96400.94920.01320.0105
LinearRow-level0.95410.94050.01180.0097
LinearConfiguration0.95420.94060.0001−0.0008
Table 10:

Grouped five-fold cross-validated train–validation gap (mean ± standard deviation). Italic: mean gap above 0.01

ModelGap LIF [–]Gap DIF [–]
Linear0.0001 ± 0.00090.0001 ± 0.0021
Polynomial0.0004 ± 0.00090.0006 ± 0.0021
Gradient boosting0.0019 ± 0.00100.0025 ± 0.0022
SVR0.0021 ± 0.00140.0037 ± 0.0026
LightGBM0.0056 ± 0.00170.0085 ± 0.0033
XGBoost0.0080 ± 0.00130.0113 ± 0.0024
Random forest0.0374 ± 0.00210.0578 ± 0.0036
Table 11:

Performance of all eleven models for LIF and DIF on the configuration-level split. Bold: selected model (DL-3 Deep). Italic: train–validation gap above 0.01. Gap = train R2 − validation R2; a negative gap indicates a higher validation than training R2

ModelTargetTrain R2 [–]Val R2 [–]Gap [–]Test R2 [–]RMSE [–]MAE [–]MAPE [%]
LinearLIF0.95160.95150.00010.95420.10190.06713.06
LinearDIF0.93730.9381−0.00080.94060.13350.09203.66
Polynomial (deg 3)LIF0.95260.95140.00120.95440.10170.06783.09
Polynomial (deg 3)DIF0.93830.93810.00020.94080.13330.09273.68
Random forestLIF0.99810.95900.03910.95890.09650.07533.39
Random forestDIF0.99760.94020.05740.93770.13670.10724.20
Gradient boostingLIF0.95490.95160.00330.95560.10030.06883.14
Gradient boostingDIF0.94100.93910.00190.94240.13150.09353.71
SVR (RBF)LIF0.95760.95130.00620.95620.09960.06833.11
SVR (RBF)DIF0.94320.93780.00550.94180.13220.09363.70
XGBoostLIF0.97150.95830.01320.96400.09030.06943.16
XGBoostDIF0.95720.94670.01050.94920.12350.09523.77
LightGBMLIF0.97330.96230.01100.96640.08730.06652.98
LightGBMDIF0.96070.95110.00960.95350.11820.09103.56
DL-1 SimpleLIF0.96660.95840.00820.96080.09420.06863.09
DL-1 SimpleDIF0.95240.94570.00660.94790.12500.09253.65
DL-2 BN+DropoutLIF0.95200.94860.00340.95210.10420.07263.34
DL-2 BN+DropoutDIF0.93620.93190.00430.93550.13920.10084.04
DL-3 DeepLIF0.97390.97050.00330.97140.08040.06482.90
DL-3 DeepDIF0.95980.95690.00290.95760.11280.08963.51
DL-4 RegularisedLIF0.96440.95610.00830.95950.09570.07053.23
DL-4 RegularisedDIF0.94970.94290.00680.94530.12810.09723.88
Figure 5:

Predicted versus FE-derived LIF for the selected LightGBM and DL-3 Deep models on the held-out test set

Figure 6:

Predicted versus FE-derived DIF for the selected LightGBM and DL-3 Deep models on the held-out test set

Figure 7:

Residual distributions for LIF and DIF predictions on the held-out test set, for LightGBM (top) and DL-3 Deep (bottom)

Figure 8:

Predicted versus FE-derived LIF for all eleven benchmarked models on a random subsample of 22 held-out test-set configurations

Figure 9:

Mean absolute error (MAE) for LIF and DIF across all eleven models, computed on the 22-configuration illustrative subsample

Figure 10:

Predicted versus FE-derived DIF for all eleven benchmarked models on a random subsample of 22 held-out test-set configurations

Figure 11:

Coefficient of determination (R2) for LIF and DIF across all eleven models, computed on the 22-configuration illustrative subsample

Table 12:

Model-based sensitivity of the LightGBM predictions on the test set: mean absolute SHAP value and permutation importance (drop in R2)

InputMean |SHAP| LIF [–]Mean |SHAP| DIF [–]Permutation ΔR2 LIF [–]Permutation ΔR2 DIF [–]
Number of stories0.3730.4291.6071.600
Bay span0.1350.1570.2310.231
Number of bays0.1110.1290.1630.164
Story height0.0050.0070.0140.015
Column removal location0.0050.0060.0130.013
Load scenario0.0000.0000.0000.000
Figure 12:

Statistical association of inputs with LIF and DIF (|Pearson r| for numerical inputs; correlation ratio η for categorical inputs)

Figure 13:

SHAP summary plot for LIF (LightGBM, test set)

Figure 14:

SHAP summary plot for DIF (LightGBM, test set)

Figure 15:

Effect of bay span on LIF and DIF across the full dataset, by column-removal location

Table 13:

Comparison of FE-derived LIF and DIF with DL-3 Deep and LightGBM predictions for six held-out configurations. Δ is the percentage deviation of the prediction from the FE-derived value; S1 to S3 denote load scenarios 1 to 3

Configuration (stories / height / span / bays / location / scenario)FE LIF [–]FE DIF [–]DL-3 LIF [–]Δ [%]DL-3 DIF [–]Δ [%]LightGBM LIF Δ [%]LightGBM DIF Δ [%]
3 / 5 m / 8 m / 6 / Corner / S31.6941.8941.712+1.11.972+4.1−0.8+1.9
2 / 3 m / 6 m / 7 / Intermediate / S21.5241.7751.492−2.11.715−3.4+16.2+6.9
10 / 4 m / 9 m / 5 / Intermediate / S12.0652.4252.184+5.72.523+4.0+4.3+1.9
16 / 3 m / 5 m / 5 / Intermediate / S32.3182.4882.158−6.92.496+0.3−6.5+0.2
28 / 3 m / 9 m / 9 / Intermediate / S13.0193.4132.843−5.83.278−4.0−4.7−2.5
28 / 3 m / 10 m / 3 / Corner / S13.2373.9663.209−0.93.702−6.7−1.8−7.3
Mean absolute deviation3.763.745.703.45
Figure 16:

Interface of the deployed web-based decision-support platform, showing the single-configuration prediction view, with input geometry, preprocessed feature vector, and classified LIF and DIF outputs

Figure 17:

Interface of the deployed web-based decision-support platform, showing the batch-prediction view with risk-band classification

DOI: https://doi.org/10.2478/cee-2027-0030 | Journal eISSN: 2199-6512 (formerly 1336-5835) | Journal ISSN: 1336-5835
Language: English
Submitted on: Sep 11, 2026
Accepted on: Oct 9, 2026
Published on: Nov 18, 2026
Published by: University of Žilina
In partnership with: Paradigm Publishing Services
Publication frequency: 4 issues per year

© 2026 Amr Mansour, Mohamed H. Serror, Sherif A. Mourad, AbdElMoniem Bayoumi, published by University of Žilina
This work is licensed under the Creative Commons Attribution 4.0 License.